Presentation
Gzswap: Mitigating GPU I/O Bottlenecks in Heterogeneous Memory Systems Through On-Device Zswap
DescriptionModern GPUs face a memory capacity wall as large-scale workloads outgrow practical device-memory limits. Scale-up designs with heterogeneous memory (HBM plus DDR and CXL) offer larger footprints, but under memory oversubscription current systems rely on Inter-memory relocation (IMR), which drives high relocation I/O traffic over off-package links and creates severe I/O bottlenecks. We propose GZswap, a hardware-managed, on-device zswap scheme that reserves a portion of GPU memory as a compressed swap region (Zpool). Instead of evicting pages to host memory, GZswap compresses and keeps them on-device, then transparently decompresses them upon reuse, converting relocation I/O traffic over off-package interconnects into on-package memory accesses. As a result, GZswap significantly reduces inter-memory I/O traffic and improves performance and energy efficiency under oversubscription, without requiring any changes to applications.
Event Type
Work in Progress
TimeMonday, July 276:43pm - 6:44pm PDT
LocationExhibit Hall